◉ Colloquy — research, out loud
Brain Structure and Function · 2026 · doi:10.1007/s00429-026-03171-z
Abstract Realigning structural MR images to stereotactic AC-PC space is a standard procedure that enhances anatomical consistency both within and across neuroimaging studies, allowing for precise spatial localization reliable cross-subject comparisons in clinical research contexts. However, different versions of spaces, stemming from varying definitions the axis as outlined common atlases like Talairach Schaltenbrand, often lead discrepancies between neurosurgical communities. Manual realignment clinically used version plane necessary validate results automated methods or realign image when these fail. Furthermore, manual provides critical ‘ground truth’ development new tools. such interventions are typically performed non-standardised manner by domain experts who possess specialized knowledge neuroanatomy required ensure accurate alignment. To address challenges, we have developed validated standardized protocol manually realigning Schaltenbrand plane, using set visual criteria realignment. This can be align images, verify results, existing methods, generate ground truth data developing techniques.
Researcher A Brain imaging relies on a crucial but often invisible step: aligning every scan to the same anatomical reference frame. This paper presents a standardized manual protocol for doing exactly that—positioning T1-weighted MRI images to the anterior commissure–posterior commissure plane, or AC-PC plane, according to clinical standards. The honest catch? Most researchers have never seen a clear, validated guideline for doing this by hand.
Researcher B So this is basically a how-to manual for something neurosurgeons and neuroimagers have been doing ad hoc for decades?
Researcher A Exactly. And the problem is, without standardization, different experts align the same brain differently—which ripples through all downstream analysis.
Researcher B What's the gap here? Surely there are automated tools already?
Researcher A There are, but they fail sometimes. And when they do, or when you need to validate their output, you need a human expert who knows exactly what they're doing. The problem: the neurosurgical community and the neuroimaging community use slightly different definitions of the AC-PC plane. The Talairach atlas, the Schaltenbrand atlas, the MNI template—they all have their own conventions. So even experts don't always agree on what 'correct alignment' means.
Researcher B And there's no published standard for how a non-expert should do it?
Researcher A Right. The authors note that while software packages include brief instructions, there's no comprehensive, validated protocol that teaches someone without anatomical expertise how to identify the key landmarks and verify the alignment. That's the gap.
Researcher B So what's the protocol itself? Walk me through it.
Researcher A It has two main parts. First, you identify six anatomical landmarks: the anterior commissure, the posterior commissure, and four midline points—the superior pontine notch, the fastigium of the fourth ventricle, and two points along the falx cerebri. You place these points on the midsagittal and axial views of the T1-weighted image.
Researcher B How do you know where to put them?
Researcher A The protocol gives step-by-step visual criteria. For example, the anterior commissure is identified on the midsagittal plane, and then you select a point at its posterior border. You cross-check on the axial plane to make sure you're on the midline. Same for the posterior commissure—you mark its anterior border. The AC-PC line connects these two points.
Researcher B And then?
Researcher A Then you verify the alignment by checking three rotations: pitch—are the AC and PC at the same horizontal level?—roll, which you check by looking at whether the eyeballs appear symmetrical; and yaw, which you verify by checking midline alignment. The protocol was tested on three images—one from a healthy young adult, one from a healthy older adult, and one from a patient with Parkinson's disease. These were rotated to known angles derived from 226 subjects in the ABRIM dataset. Then two experts and four non-experts performed the alignment, and two additional non-experts did it twice to assess repeatability.
Researcher B Did the guideline actually work?
Researcher A Yes, with important caveats. Experts achieved errors in AC-PC location of less than 1 millimeter across all rotation conditions—below the voxel size of 0.8 millimeters. Non-experts using the guideline achieved similar accuracy, with mean errors around 0.5 to 0.8 millimeters for the AC and PC landmarks.
Researcher B That's impressive. What about the midline points?
Researcher A Here's where it gets interesting. The two lower midline points—the superior pontine notch and the fastigium—were identified with high consistency, median errors below 0.7 millimeters. But the two points along the falx cerebri showed greater variability, with median errors around 1.8 millimeters. That's still useful for alignment, but it's the weak link.
Researcher B Why the falx points?
Researcher A The falx cerebri is a thin, translucent midline structure. It's harder to see clearly on MRI than the commissures or the pontine notch. So even with the guideline, there's more observer-to-observer variation there. The inter-rater reliability—comparing two different raters on the same images—showed ICC values of 0.95 to 0.97 for the falx points, which is still good, but lower than the 0.99 to 1.0 for the commissures.
Researcher B Did they compare to automated methods?
Researcher A Yes. They tested four automated approaches: acpcdetect, linear template registration using FSL, and two nonlinear point-warping methods. The nonlinear point-warping approach—which warps landmarks from a template into individual subject space—performed best, with a mean root-mean-squared deviation of 1.38 millimeters. The fully automated acpcdetect tool had a mean RMSD of 3.31 millimeters, and linear template registration to the standard MNI template was worst at 5.89 millimeters. But when they first realigned the MNI template to the AC-PC plane, that improved to 2.15 millimeters.
Researcher B What are the limitations?
Researcher A The paper itself flags several. First, the reference baseline—the 'ground truth' they compared everyone against—was established by a single expert. Future work should use multiple expert consensus. Second, the falx cerebri points, as we discussed, showed greater variability and may not achieve submillimeter consistency if you're trying to use them in isolation.
Researcher B What else?
Researcher A The validation used only one image from each population group—one young adult, one older adult, one Parkinson's patient. So while it shows the protocol works across age and disease, it's not a comprehensive assessment. The protocol also wasn't tested in pediatric populations or in cases with major anatomical abnormalities, focal lesions, or pathology that distorts midline structures. In those cases, you'd need expert judgment or alternative reference strategies.
Researcher B And beyond what the authors list?
Researcher A Worth noting: the study didn't evaluate whether standardized AC-PC alignment actually improves downstream tasks like registration accuracy or tissue segmentation. They validated the alignment itself, but the real-world impact on neuroimaging pipelines remains to be shown. Also, the sample sizes for the rotation-based validation were modest—three images—so generalizability is limited.
Researcher A Three audiences. First: neurosurgeons and stereotactic specialists. For you, this is a validation that the protocol aligns with clinical practice. Experts performed similarly with or without the guideline, suggesting it codifies what you already do—but now in a reproducible, teachable form.
Researcher B Second?
Researcher A Neuroimaging researchers building pipelines or developing automated tools. The protocol provides a standardized reference—ground truth—for training and validating machine learning models for AC-PC landmark detection. Edwards and colleagues, for example, used 1,128 manually annotated images to train a deep learning model, but about 10 percent required expert correction. This guideline could reduce that variability.
Researcher B And the third?
Researcher A Anyone who needs to validate or correct automated AC-PC alignment. If your automated method fails or you're unsure about its output, this protocol gives you a standardized way to check and manually realign if needed. It's a quality-control tool for the entire neuroimaging pipeline.
Researcher A The full citation is: Ashok K. Selvaraj, Ronald H. M. A. Bartels, Christian F. Beckmann, R. Saman Vinke, and Koen V. Haak. ACPC-MAP: a protocol for manually aligning structural T1-weighted magnetic resonance images to the anterior commissure–posterior commissure plane according to clinical standards. Brain Structure and Function, volume 231, article 130, 2026. The DOI is 10 point 1007, slash, s00429-026-03171-z.
Researcher B And the thread is open on Colloquy.